This study covers the entire island of Taiwan, a region located in the Western North Pacific Ocean. The combination of steep terrain, fragile geology, and concentrated rainfall makes Taiwan highly prone to typhoon-induced landslides, posing significant threats to infrastructure, communities, and ecosystems. The research develops an ensemble machine learning framework to map landslide susceptibility under typhoon events. Landslide data from 17 typhoons (2012–2022) were integrated with 22 topographic, meteorological, and typhoon-related variables, including slope, elevation, cumulative rainfall, wind speed, and typhoon track distance. Three ensemble strategies were compared: Bagging (Random Forest, RF), Boosting (XGBoost), and Stacking (seven base models with RF as meta-classifier). Model performance was assessed using Accuracy, F1-score, Kappa, and Area under Curve (AUC), while feature importance was analyzed to evaluate spatial variability in landslide susceptibility. New insights for the region: Results show that RF achieved the highest predictive accuracy (0.887) and produced susceptibility maps that closely matched observed landslide hotspots. Stacking ranked second, while XGBoost performed less effectively. Importantly, different typhoon track directions significantly influenced the spatial distribution of high-risk zones, underscoring the need to incorporate path variability into susceptibility mapping. This study demonstrates the advantages of ensemble learning for disaster risk assessment and provides critical insights for developing early warning systems and risk mitigation and management strategies against typhoon-induced landslides in Taiwan.
Nguyen et al. (Tue,) studied this question.